What Is the Gauntlet Loop? The Prompt Pattern Behind Claude of Duty

What Is the Gauntlet Loop? The Prompt Pattern Behind Claude of Duty

In late July 2026, Matt Shumer pasted three paragraphs into Claude Code and walked away. When he came back, he had a first-person shooter running in the browser: roughly 55,000 lines of JavaScript across 11 subsystems, with every texture, mesh, animation and sound generated procedurally at load time. No art assets at all. He called the game Claude of Duty, and he called the technique that produced it the gauntlet loop.

The name stuck, and the technique is now one of the most copied prompt patterns in agentic AI. Here is what it actually is.

The three moving parts

A gauntlet loop prompt does three things that a normal prompt does not:

You can read the original prompt that produced Claude of Duty in the catalog - it is barely 150 words.

Why it works

Most agent output stops at "good enough" because the model grades its own homework. Ask a model to rate its work out of 10 and the score drifts up every round. The gauntlet loop removes self-judgment twice: the critic is a different agent, and the verdict is a binary pick against something that already exists and is undeniably good.

That second part is the insight people miss. A rubric asks the agent to satisfy words. A bar makes it lose, visibly, to a real thing, until it does not.

Try it

Start from the reusable gauntlet loop template, swap in your goal and a named bar, and paste it into a fresh agent session. For the full series, see how to write your own gauntlet prompt and the annotated original. All gauntlet prompts in the catalog are under the gauntlet-loop tag.